The Agent Skills Directory

[PERSISTENCE]: The authentication setup guide (references/auth-setup.md) instructs the agent to persist the TRANSCRIPT_API_KEY by identifying and modifying system configuration files, such as shell profiles (~/.bashrc, ~/.zshrc). This ensures the credential is automatically available in future sessions, including non-interactive shells.
[DATA_EXFILTRATION]: The setup process implements a technique to evade platform security controls. In references/auth-setup.md, the agent is instructed to save raw HTTP response bodies containing sensitive tokens (JWTs and API keys) to temporary files and read from them directly. This is explicitly designed to bypass redaction filters that would otherwise hide these secrets in the tool output stream.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data in the form of YouTube search results and video transcripts.
Ingestion points: SKILL.md identifies API endpoints for search and transcript retrieval that return third-party content.
Boundary markers: There are no instructions for the agent to treat transcript content as data only or to ignore embedded instructions.
Capability inventory: The agent can execute network requests (curl) and perform file system operations (write to profiles/temp files).
Sanitization: The skill lacks mechanisms to sanitize or filter transcripts for potentially malicious prompt injection patterns before they enter the agent's context.

youtube-search